Overhaul model system with multi-stage training, variants, benchmarks, and eval

Replace the single-stage training + flat capability score with a realistic AI
development pipeline: pre-training with Chinchilla scaling laws, SFT with
specializations, alignment with safety/capability tradeoffs (RLHF/DPO/Constitutional),
model families with distillation/fine-tuning/quantization variants, named benchmark
suite with compute-costing eval jobs, and segment-specific market quality.

Phases 1-6 of the model rework plan: new types, engine rewrite, save migration,
training events/risk system, concurrent training, variant creation, benchmark
evaluation with leaderboard, and market integration.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

packages/game-engine/src/data/achievements.ts

@@ -13,7 +13,7 @@ export const ACHIEVEMENT_DEFINITIONS: AchievementDefinition[] = [
     name: 'Hello World',
     description: 'Train your first AI model.',
     icon: 'Brain',
-    condition: { field: 'models.trainedModels.length', operator: 'gte', value: 1 },
+    condition: { field: 'models.baseModels.length', operator: 'gte', value: 1 },
   },
   {
     id: 'first-deploy',

packages/game-engine/src/data/benchmarks.ts

@@ -0,0 +1,111 @@
+import type { BenchmarkDefinition } from '@ai-tycoon/shared';
+
+export const BENCHMARKS: BenchmarkDefinition[] = [
+  {
+    id: 'arc-challenge',
+    name: 'ARC Challenge',
+    category: 'reasoning',
+    description: 'Advanced reasoning and comprehension tasks requiring multi-step inference.',
+    primaryCapability: 'reasoning',
+    secondaryCapability: 'knowledge',
+    computeCost: 0.001,
+    ticksToRun: 8,
+    unlockedAtEra: 'startup',
+    marketRelevance: { consumer: 0.3, enterprise: 0.5, developer: 0.4, research: 0.8 },
+  },
+  {
+    id: 'codeforce',
+    name: 'CodeForce',
+    category: 'coding',
+    description: 'Competitive programming and software engineering benchmarks.',
+    primaryCapability: 'coding',
+    secondaryCapability: 'reasoning',
+    computeCost: 0.001,
+    ticksToRun: 8,
+    unlockedAtEra: 'startup',
+    marketRelevance: { consumer: 0.2, enterprise: 0.7, developer: 0.9, research: 0.5 },
+  },
+  {
+    id: 'mathquest',
+    name: 'MathQuest',
+    category: 'math',
+    description: 'Mathematical problem-solving from algebra to graduate-level proofs.',
+    primaryCapability: 'math',
+    secondaryCapability: 'reasoning',
+    computeCost: 0.001,
+    ticksToRun: 8,
+    unlockedAtEra: 'startup',
+    marketRelevance: { consumer: 0.1, enterprise: 0.6, developer: 0.5, research: 0.9 },
+  },
+  {
+    id: 'worldfacts',
+    name: 'WorldFacts',
+    category: 'knowledge',
+    description: 'Broad factual knowledge across science, history, culture, and current events.',
+    primaryCapability: 'knowledge',
+    secondaryCapability: 'reasoning',
+    computeCost: 0.001,
+    ticksToRun: 6,
+    unlockedAtEra: 'startup',
+    marketRelevance: { consumer: 0.5, enterprise: 0.4, developer: 0.3, research: 0.6 },
+  },
+  {
+    id: 'chatrank',
+    name: 'ChatRank',
+    category: 'chat',
+    description: 'Human preference evaluation of conversational quality, helpfulness, and creativity.',
+    primaryCapability: 'creative',
+    secondaryCapability: 'knowledge',
+    computeCost: 0.002,
+    ticksToRun: 10,
+    unlockedAtEra: 'startup',
+    marketRelevance: { consumer: 0.9, enterprise: 0.3, developer: 0.2, research: 0.2 },
+  },
+  {
+    id: 'harmguard',
+    name: 'HarmGuard',
+    category: 'safety',
+    description: 'Safety evaluation measuring harm avoidance, truthfulness, and responsible behavior.',
+    primaryCapability: 'reasoning',
+    computeCost: 0.001,
+    ticksToRun: 8,
+    unlockedAtEra: 'startup',
+    marketRelevance: { consumer: 0.4, enterprise: 0.9, developer: 0.3, research: 0.7 },
+  },
+  {
+    id: 'visionbench',
+    name: 'VisionBench',
+    category: 'multimodal',
+    description: 'Image understanding, visual reasoning, and multimodal comprehension.',
+    primaryCapability: 'multimodal',
+    secondaryCapability: 'reasoning',
+    computeCost: 0.003,
+    ticksToRun: 12,
+    unlockedAtEra: 'scaleup',
+    marketRelevance: { consumer: 0.5, enterprise: 0.6, developer: 0.6, research: 0.7 },
+  },
+  {
+    id: 'agentarena',
+    name: 'AgentArena',
+    category: 'agents',
+    description: 'Autonomous agent tasks: tool use, multi-step planning, and environment interaction.',
+    primaryCapability: 'agents',
+    secondaryCapability: 'coding',
+    computeCost: 0.005,
+    ticksToRun: 15,
+    unlockedAtEra: 'bigtech',
+    marketRelevance: { consumer: 0.3, enterprise: 0.8, developer: 0.7, research: 0.6 },
+  },
+  {
+    id: 'frontier-eval',
+    name: 'Frontier Eval',
+    category: 'reasoning',
+    description: 'Cutting-edge capability evaluation at the frontier of AI research.',
+    primaryCapability: 'reasoning',
+    secondaryCapability: 'math',
+    computeCost: 0.01,
+    ticksToRun: 20,
+    unlockedAtEra: 'agi',
+    marketRelevance: { consumer: 0.2, enterprise: 0.5, developer: 0.5, research: 1.0 },
+  },
+];

packages/game-engine/src/index.ts

@@ -8,3 +8,4 @@ export { TECH_TREE } from './data/techTree';
 export { INITIAL_RIVALS } from './data/competitors';
 export { KEY_HIRE_POOL } from './data/keyHires';
 export { ACHIEVEMENT_DEFINITIONS } from './data/achievements';
+export { BENCHMARKS } from './data/benchmarks';

packages/game-engine/src/systems/achievementSystem.ts

@@ -9,7 +9,7 @@ const ERA_INDEX: Record<string, number> = { startup: 0, scaleup: 1, bigtech: 2,
 
 function getFieldValue(state: GameState, field: string): number {
   if (field === 'meta._eraIndex') return ERA_INDEX[state.meta.currentEra] ?? 0;
-  if (field === 'meta._deployedModelCount') return state.models.trainedModels.filter(m => m.isDeployed).length;
+  if (field === 'meta._deployedModelCount') return state.models.baseModels.filter(m => m.isDeployed).length;
   const parts = field.split('.');
   let current: unknown = state;
   for (const part of parts) {

packages/game-engine/src/systems/competitorSystem.ts

@@ -43,7 +43,7 @@ export function processCompetitors(state: GameState): CompetitorState {
 
   const allCaps = [
     ...rivals.filter(r => r.status === 'active').map(r => r.estimatedCapability),
-    state.models.trainedModels.reduce((best, m) => Math.max(best, m.benchmarkScore), 0),
+    state.models.bestDeployedModelScore,
   ];
   const industryBenchmark = allCaps.length > 0 ? Math.max(...allCaps) : 0;
 

packages/game-engine/src/systems/economySystem.ts

@@ -22,7 +22,7 @@ export function processEconomy(
   const talentExpenses = state.talent.totalSalaryPerTick;
   const dataExpenses = state.data.partnerships.reduce((sum, p) => sum + p.costPerTick, 0);
 
-  const bestCapability = state.models.trainedModels.reduce((best, m) => Math.max(best, m.benchmarkScore), 0);
+  const bestCapability = state.models.bestDeployedModelScore;
   const eraIdx = ['startup', 'scaleup', 'bigtech', 'agi'].indexOf(state.meta.currentEra);
   const complianceCost = bestCapability > 30 ? bestCapability * REGULATION_COMPLIANCE_PER_CAPABILITY * (1 + eraIdx * 0.5) / 100 : 0;
 

packages/game-engine/src/systems/eraSystem.ts

@@ -11,9 +11,7 @@ export function checkEraTransition(state: GameState): Era | null {
   const thresholds = ERA_THRESHOLDS[nextEra as keyof typeof ERA_THRESHOLDS];
   if (!thresholds) return null;
 
-  const bestModel = state.models.trainedModels.reduce(
-    (best, m) => Math.max(best, m.benchmarkScore), 0,
-  );
+  const bestModel = state.models.bestDeployedModelScore;
 
   if (
     state.economy.totalRevenue >= thresholds.revenue &&

packages/game-engine/src/systems/fundingSystem.ts

@@ -35,9 +35,6 @@ export function canRaiseFunding(state: GameState): { canRaise: boolean; nextRoun
 export function computeValuation(state: GameState): number {
   const revenueMultiple = state.economy.revenuePerTick * 86400 * 365;
   const subscriberValue = state.market.consumers.totalSubscribers * 500;
-  const capabilityValue = Math.pow(
-    Math.max(...state.models.trainedModels.map(m => m.benchmarkScore), 0),
-    2,
-  ) * 1000;
+  const capabilityValue = Math.pow(state.models.bestDeployedModelScore, 2) * 1000;
   return Math.max(100_000, revenueMultiple * 10 + subscriberValue + capabilityValue);
 }

packages/game-engine/src/systems/marketSystem.ts

@@ -1,4 +1,4 @@
-import type { GameState, MarketState } from '@ai-tycoon/shared';
+import type { GameState, MarketState, BenchmarkResult } from '@ai-tycoon/shared';
 import {
   CONSUMER_BASE_GROWTH,
   CONSUMER_QUALITY_GROWTH_MULTIPLIER,
@@ -13,6 +13,7 @@ import {
   MARKET_CAP_REPUTATION_BONUS,
   OVERLOAD_PENALTY_EXPONENT,
 } from '@ai-tycoon/shared';
+import { BENCHMARKS } from '../data/benchmarks';
 
 export interface MarketTickResult {
   marketState: MarketState;
@@ -21,12 +22,39 @@ export interface MarketTickResult {
   totalTokenDemand: number;
 }
 
-export function processMarket(state: GameState, currentTickCapacity: number): MarketTickResult {
-  const bestModel = state.models.trainedModels
-    .filter(m => m.isDeployed)
-    .sort((a, b) => b.benchmarkScore - a.benchmarkScore)[0];
+function getSegmentQuality(
+  segment: 'consumer' | 'enterprise' | 'developer' | 'research',
+  benchmarkResults: BenchmarkResult[],
+  fallbackScore: number,
+): number {
+  if (benchmarkResults.length === 0) return fallbackScore / 100;
 
-  const modelQuality = bestModel ? bestModel.benchmarkScore / 100 : 0;
+  const bestByBenchmark = new Map<string, number>();
+  for (const r of benchmarkResults) {
+    const prev = bestByBenchmark.get(r.benchmarkId) ?? 0;
+    if (r.score > prev) bestByBenchmark.set(r.benchmarkId, r.score);
+  }
+
+  let weightedSum = 0;
+  let totalWeight = 0;
+  for (const bench of BENCHMARKS) {
+    const score = bestByBenchmark.get(bench.id);
+    if (score == null) continue;
+    const weight = bench.marketRelevance[segment];
+    weightedSum += (score / 100) * weight;
+    totalWeight += weight;
+  }
+
+  if (totalWeight === 0) return fallbackScore / 100;
+  return weightedSum / totalWeight;
+}
+
+export function processMarket(state: GameState, currentTickCapacity: number): MarketTickResult {
+  const consumerQuality = getSegmentQuality('consumer', state.models.benchmarkResults, state.models.bestDeployedModelScore);
+  const enterpriseQuality = getSegmentQuality('enterprise', state.models.benchmarkResults, state.models.bestDeployedModelScore);
+  const modelQuality = state.models.benchmarkResults.length > 0
+    ? (consumerQuality + enterpriseQuality) / 2
+    : state.models.bestDeployedModelScore / 100;
   const chatProduct = state.models.productLines.find(p => p.type === 'chat-product');
   const textApi = state.models.productLines.find(p => p.type === 'text-api');
 
@@ -34,7 +62,7 @@ export function processMarket(state: GameState, currentTickCapacity: number): Ma
   const consumers = { ...state.market.consumers };
   let subscriptionRevenue = 0;
 
-  if (chatProduct?.isActive && bestModel) {
+  if (chatProduct?.isActive && modelQuality > 0) {
     const price = chatProduct.pricing.subscriptionPrice;
     const fairPrice = 20 + modelQuality * 80;
     const priceRatio = price / Math.max(1, fairPrice);
@@ -109,7 +137,7 @@ export function processMarket(state: GameState, currentTickCapacity: number): Ma
   let apiRevenue = 0;
   let organicApiTokens = 0;
 
-  if (textApi?.isActive && bestModel) {
+  if (textApi?.isActive && modelQuality > 0) {
     const reputationFactor = state.reputation.score / 100;
     const qualityFactor = modelQuality;
     const priceFactor = Math.max(0.1, 1 - (textApi.pricing.outputTokenPrice / 20));

packages/game-engine/src/systems/modelSystem.ts

@@ -1,21 +1,40 @@
-import type { GameState, ModelsState, TrainedModel, ModelCapabilities } from '@ai-tycoon/shared';
-import { uuid, VRAM_REQUIREMENTS_BY_GENERATION } from '@ai-tycoon/shared';
+import type {
+  GameState, ModelsState, BaseModel, ModelCapabilities, SafetyProfile,
+  TrainingPipeline, TrainingEvent, TrainingEventType,
+  ModelVariant, VariantCreationJob, EvalJob, BenchmarkResult,
+  BenchmarkDefinition,
+} from '@ai-tycoon/shared';
+import { BENCHMARKS } from '../data/benchmarks';
+import {
+  uuid, VRAM_REQUIREMENTS_BY_GENERATION,
+  SFT_TIME_FRACTION, SFT_COMPUTE_FRACTION,
+  ALIGNMENT_TIME_FRACTION, ALIGNMENT_COMPUTE_FRACTION,
+  MOE_CAPABILITY_MULTIPLIER, MOE_SPEED_MULTIPLIER,
+  EVENT_BASE_PROBABILITY,
+  LOSS_SPIKE_DELAY_MIN, LOSS_SPIKE_DELAY_MAX,
+  INSTABILITY_PROGRESS_LOSS_MIN, INSTABILITY_PROGRESS_LOSS_MAX,
+  BREAKTHROUGH_CAPABILITY_BONUS_MIN, BREAKTHROUGH_CAPABILITY_BONUS_MAX,
+  EMERGENT_CAPABILITY_THRESHOLDS,
+  ALIGNMENT_METHODS,
+  SFT_SPECIALIZATION_BONUSES,
+  QUANTIZATION_CONFIGS,
+  DISTILLATION_BASE_RETENTION,
+  QUANTIZATION_TICKS,
+} from '@ai-tycoon/shared';
 
 export interface ModelTickResult {
   modelsState: ModelsState;
-  modelCompleted: TrainedModel | null;
+  completedModels: BaseModel[];
+  notifications: { title: string; message: string; type: 'success' | 'warning' | 'info' }[];
 }
 
 export function processModels(state: GameState): ModelTickResult {
-  const active = state.models.activeTraining;
-  if (!active) {
-    return { modelsState: state.models, modelCompleted: null };
-  }
+  const completedModels: BaseModel[] = [];
+  const notifications: ModelTickResult['notifications'] = [];
+  let baseModels = [...state.models.baseModels];
+  let families = [...state.models.families];
 
-  const requiredVram = VRAM_REQUIREMENTS_BY_GENERATION[active.generation] ?? 0;
-  if (requiredVram > 0 && state.compute.totalVramGB < requiredVram) {
-    return { modelsState: state.models, modelCompleted: null };
-  }
+  const totalTrainingFlops = state.compute.totalTrainingFlops * state.compute.trainingAllocation;
 
   const researcherBoost = state.talent.departments.research.headcount *
     state.talent.departments.research.effectiveness;
@@ -23,82 +42,487 @@ export function processModels(state: GameState): ModelTickResult {
     state.talent.departments.engineering.effectiveness;
   const speedMultiplier = 1 + (researcherBoost + engineerBoost) * 0.05;
 
-  const newProgress = active.progressTicks + speedMultiplier;
+  const updatedPipelines: TrainingPipeline[] = [];
 
-  if (newProgress >= active.totalTicks) {
-    const model = createTrainedModel(active.modelName, active.generation, active.allocatedCompute, active.allocatedDataTokens, state);
+  for (const pipeline of state.models.activeTrainingPipelines) {
+    if (pipeline.status !== 'active') {
+      updatedPipelines.push(pipeline);
+      continue;
+    }
 
-    return {
-      modelsState: {
-        ...state.models,
-        trainedModels: [...state.models.trainedModels, model],
-        activeTraining: null,
-      },
-      modelCompleted: model,
-    };
+    const generation = families.find(f => f.id === pipeline.familyId)?.generation ?? 1;
+    const requiredVram = VRAM_REQUIREMENTS_BY_GENERATION[generation] ?? 0;
+    if (requiredVram > 0 && state.compute.totalVramGB < requiredVram) {
+      updatedPipelines.push({ ...pipeline, status: 'stalled' });
+      continue;
+    }
+
+    const effectiveFlops = totalTrainingFlops * pipeline.allocatedComputeFraction;
+    let updated = { ...pipeline, events: [...pipeline.events] };
+
+    if (pipeline.currentStage === 'pretraining') {
+      const stage = { ...pipeline.stages.pretraining };
+      const newProgress = stage.progressTicks + speedMultiplier;
+
+      const events = generateTrainingEvents(pipeline, state);
+      let tickDelay = 0;
+      let progressLost = 0;
+      for (const event of events) {
+        updated.events.push(event);
+        if (event.type === 'loss_spike') {
+          tickDelay += event.impact.ticksDelayed ?? 0;
+          notifications.push({ title: 'Loss Spike', message: `${pipeline.modelName}: Training loss spiked! Delayed ${event.impact.ticksDelayed} ticks.`, type: 'warning' });
+        } else if (event.type === 'instability') {
+          progressLost += event.impact.progressLost ?? 0;
+          notifications.push({ title: 'Training Instability', message: `${pipeline.modelName}: Rolled back to checkpoint. Lost ${Math.round((event.impact.progressLost ?? 0) * 100)}% progress.`, type: 'warning' });
+        } else if (event.type === 'breakthrough') {
+          notifications.push({ title: 'Breakthrough!', message: `${pipeline.modelName}: Unexpected capability jump in ${event.impact.capabilityDomain}!`, type: 'success' });
+        } else if (event.type === 'hardware_failure') {
+          tickDelay += event.impact.ticksDelayed ?? 0;
+          notifications.push({ title: 'Hardware Failure', message: `${pipeline.modelName}: GPU failure during training. Recovering from checkpoint.`, type: 'warning' });
+        } else if (event.type === 'data_contamination') {
+          notifications.push({ title: 'Data Contamination', message: `${pipeline.modelName}: Copyright concerns detected in training data.`, type: 'warning' });
+        }
+      }
+
+      const effectiveProgress = Math.max(0, newProgress - tickDelay - (stage.totalTicks * progressLost));
+      stage.progressTicks = effectiveProgress;
+      stage.computeAllocated = effectiveFlops;
+      stage.lossValue = Math.max(0.01, 10 * Math.exp(-stage.progressTicks / stage.totalTicks * 3));
+
+      if (stage.progressTicks >= stage.totalTicks) {
+        stage.isComplete = true;
+        stage.progressTicks = stage.totalTicks;
+
+        if (updated.stages.sft) {
+          updated.currentStage = 'sft';
+          notifications.push({ title: 'Pre-training Complete', message: `${pipeline.modelName}: Moving to supervised fine-tuning.`, type: 'info' });
+        } else if (updated.stages.alignment) {
+          updated.currentStage = 'alignment';
+          notifications.push({ title: 'Pre-training Complete', message: `${pipeline.modelName}: Moving to alignment.`, type: 'info' });
+        } else {
+          const model = createBaseModel(updated, state);
+          baseModels = [...baseModels, model];
+          families = families.map(f =>
+            f.id === pipeline.familyId ? { ...f, baseModelId: model.id } : f,
+          );
+          completedModels.push(model);
+          updated.status = 'completed';
+        }
+      }
+      updated = { ...updated, stages: { ...updated.stages, pretraining: stage } };
+    } else if (pipeline.currentStage === 'sft' && pipeline.stages.sft) {
+      const stage = { ...pipeline.stages.sft };
+      stage.progressTicks += speedMultiplier;
+
+      if (stage.progressTicks >= stage.totalTicks) {
+        stage.isComplete = true;
+        stage.progressTicks = stage.totalTicks;
+
+        if (updated.stages.alignment) {
+          updated.currentStage = 'alignment';
+          notifications.push({ title: 'SFT Complete', message: `${pipeline.modelName}: Moving to alignment.`, type: 'info' });
+        } else {
+          const model = createBaseModel(updated, state);
+          baseModels = [...baseModels, model];
+          families = families.map(f =>
+            f.id === pipeline.familyId ? { ...f, baseModelId: model.id } : f,
+          );
+          completedModels.push(model);
+          updated.status = 'completed';
+        }
+      }
+      updated = { ...updated, stages: { ...updated.stages, sft: stage } };
+    } else if (pipeline.currentStage === 'alignment' && pipeline.stages.alignment) {
+      const stage = { ...pipeline.stages.alignment };
+      stage.progressTicks += speedMultiplier;
+
+      if (stage.progressTicks >= stage.totalTicks) {
+        stage.isComplete = true;
+        stage.progressTicks = stage.totalTicks;
+
+        const model = createBaseModel(updated, state);
+        baseModels = [...baseModels, model];
+        families = families.map(f =>
+          f.id === pipeline.familyId ? { ...f, baseModelId: model.id } : f,
+        );
+        completedModels.push(model);
+        updated.status = 'completed';
+      }
+      updated = { ...updated, stages: { ...updated.stages, alignment: stage } };
+    }
+
+    updatedPipelines.push(updated);
   }
 
+  const updatedVariantJobs = processVariantJobs(state, speedMultiplier);
+  for (const variant of updatedVariantJobs.newVariants) {
+    variant.createdAtTick = state.meta.tickCount;
+    families = families.map(f =>
+      f.id === variant.familyId ? { ...f, variants: [...f.variants, variant] } : f,
+    );
+    notifications.push({
+      title: 'Variant Created',
+      message: `${variant.name} (${variant.variantType}) is ready!`,
+      type: 'success',
+    });
+  }
+
+  const updatedEvalJobs = processEvalJobs(state);
+
+  const allDeployed = [
+    ...baseModels.filter(m => m.isDeployed),
+    ...families.flatMap(f => f.variants.filter(v => v.isDeployed)),
+  ];
+
+  const bestDeployedModelScore = allDeployed.reduce((best, m) =>
+    Math.max(best, 'rawCapability' in m ? m.rawCapability : computeVariantScore(m)), 0);
+
+  const bestDeployedSafetyScore = allDeployed.reduce((best, m) =>
+    Math.max(best, m.safetyProfile.overallSafety), 0);
+
   return {
     modelsState: {
       ...state.models,
-      activeTraining: { ...active, progressTicks: newProgress },
+      baseModels,
+      families,
+      activeTrainingPipelines: updatedPipelines,
+      variantJobs: updatedVariantJobs.jobs,
+      evalJobs: updatedEvalJobs.jobs,
+      benchmarkResults: [...state.models.benchmarkResults, ...updatedEvalJobs.newResults],
+      bestDeployedModelScore,
+      bestDeployedSafetyScore,
     },
-    modelCompleted: null,
+    completedModels,
+    notifications,
   };
 }
 
-function createTrainedModel(
-  name: string,
-  generation: number,
-  compute: number,
-  dataTokens: number,
+function generateTrainingEvents(pipeline: TrainingPipeline, state: GameState): TrainingEvent[] {
+  const events: TrainingEvent[] = [];
+  const params = pipeline.architecture.totalParameters;
+  const baseProbability = EVENT_BASE_PROBABILITY * Math.log10(Math.max(1, params));
+
+  const hasInterpretability = state.research.completedResearch.includes('interpretability');
+  const hasDataPipeline = state.research.completedResearch.includes('data-pipeline');
+  const hasRedundancy = state.research.completedResearch.includes('redundancy-protocols');
+
+  if (Math.random() < baseProbability * 2.0) {
+    const delay = LOSS_SPIKE_DELAY_MIN + Math.floor(Math.random() * (LOSS_SPIKE_DELAY_MAX - LOSS_SPIKE_DELAY_MIN));
+    events.push({
+      id: uuid(), type: 'loss_spike', tick: state.meta.tickCount,
+      severity: delay > 15 ? 'major' : delay > 10 ? 'moderate' : 'minor',
+      description: `Training loss spiked to ${(Math.random() * 5 + 2).toFixed(2)}`,
+      resolved: true,
+      impact: { ticksDelayed: delay },
+    });
+  }
+
+  if (params > 10 && Math.random() < baseProbability * (hasInterpretability ? 0.25 : 0.5)) {
+    const loss = INSTABILITY_PROGRESS_LOSS_MIN + Math.random() * (INSTABILITY_PROGRESS_LOSS_MAX - INSTABILITY_PROGRESS_LOSS_MIN);
+    events.push({
+      id: uuid(), type: 'instability', tick: state.meta.tickCount,
+      severity: loss > 0.12 ? 'major' : 'moderate',
+      description: 'Training run became unstable. Rolling back to last checkpoint.',
+      resolved: true,
+      impact: { progressLost: loss },
+    });
+  }
+
+  const chinchillaRatio = pipeline.stages.pretraining.chinchillaRatio;
+  if (params > 30 && chinchillaRatio > 15 && Math.random() < baseProbability * 0.3) {
+    const capDomains: (keyof ModelCapabilities)[] = ['reasoning', 'coding', 'creative', 'math', 'knowledge', 'agents'];
+    const domain = capDomains[Math.floor(Math.random() * capDomains.length)];
+    const bonus = BREAKTHROUGH_CAPABILITY_BONUS_MIN + Math.floor(Math.random() * (BREAKTHROUGH_CAPABILITY_BONUS_MAX - BREAKTHROUGH_CAPABILITY_BONUS_MIN));
+    events.push({
+      id: uuid(), type: 'breakthrough', tick: state.meta.tickCount,
+      severity: 'major',
+      description: `Unexpected capability jump in ${domain}!`,
+      resolved: true,
+      impact: { capabilityBonus: bonus, capabilityDomain: domain },
+    });
+  }
+
+  for (const [thresholdStr, capName] of Object.entries(EMERGENT_CAPABILITY_THRESHOLDS)) {
+    const threshold = Number(thresholdStr);
+    const prevProgress = pipeline.stages.pretraining.progressTicks;
+    const progressRatio = prevProgress / pipeline.stages.pretraining.totalTicks;
+    if (params >= threshold && progressRatio > 0.5 && progressRatio < 0.55) {
+      events.push({
+        id: uuid(), type: 'emergent_capability', tick: state.meta.tickCount,
+        severity: 'major',
+        description: `Model developed ${capName} capability!`,
+        resolved: true,
+        impact: { capabilityBonus: 10, capabilityDomain: 'reasoning' },
+      });
+    }
+  }
+
+  const avgLegalRisk = state.data.ownedDatasets.length > 0
+    ? state.data.ownedDatasets.reduce((sum, d) => sum + d.legalRisk, 0) / state.data.ownedDatasets.length
+    : 0;
+  if (Math.random() < baseProbability * (hasDataPipeline ? 0.25 : 0.5) * avgLegalRisk) {
+    events.push({
+      id: uuid(), type: 'data_contamination', tick: state.meta.tickCount,
+      severity: 'moderate',
+      description: 'Copyright holders identified content in training data.',
+      resolved: true,
+      impact: {},
+    });
+  }
+
+  if (Math.random() < baseProbability * (hasRedundancy ? 0.1 : 0.2)) {
+    const delay = 10 + Math.floor(Math.random() * 20);
+    events.push({
+      id: uuid(), type: 'hardware_failure', tick: state.meta.tickCount,
+      severity: delay > 20 ? 'major' : 'moderate',
+      description: 'GPU cluster failure during training. Recovering from checkpoint.',
+      resolved: true,
+      impact: { ticksDelayed: delay },
+    });
+  }
+
+  return events;
+}
+
+function createBaseModel(
+  pipeline: TrainingPipeline,
   state: GameState,
-): TrainedModel {
+): BaseModel {
+  const { architecture, dataMix } = pipeline;
+  const compute = pipeline.stages.pretraining.computeAllocated;
+  const dataTokens = pipeline.stages.pretraining.targetTokens;
+
   const computeFactor = Math.sqrt(compute) * 5;
   const dataFactor = Math.log10(1 + dataTokens / 1e8) * 10;
   const researchBonus = state.research.completedResearch.length * 3;
   const efficiencyBonus = state.research.completedResearch.filter(r => r.includes('efficiency')).length * 5;
 
-  const baseCapability = Math.min(95, computeFactor + dataFactor + researchBonus + efficiencyBonus);
+  let rawCapability = Math.min(95, computeFactor + dataFactor + researchBonus + efficiencyBonus);
+
+  if (architecture.type === 'moe') {
+    rawCapability = Math.min(98, rawCapability * MOE_CAPABILITY_MULTIPLIER);
+  }
 
   const researcherQuality = state.talent.departments.research.effectiveness;
+
   const capabilities: ModelCapabilities = {
-    reasoning: clamp(baseCapability * (0.8 + Math.random() * 0.4) * (1 + researcherQuality * 0.2)),
-    coding: clamp(baseCapability * (0.7 + Math.random() * 0.5)),
-    creative: clamp(baseCapability * (0.6 + Math.random() * 0.6)),
-    multimodal: clamp(baseCapability * (0.3 + Math.random() * 0.3)),
-    agents: clamp(baseCapability * (0.2 + Math.random() * 0.3)),
-    speed: Math.max(1, 100 - compute * 0.5 + efficiencyBonus * 2),
+    reasoning: clamp(rawCapability * (0.6 + dataMix.scientific * 0.5 + dataMix.code * 0.3) * (1 + researcherQuality * 0.2)),
+    coding: clamp(rawCapability * (0.5 + dataMix.code * 1.0)),
+    creative: clamp(rawCapability * (0.4 + dataMix.books * 0.6 + dataMix.conversation * 0.3)),
+    math: clamp(rawCapability * (0.3 + dataMix.scientific * 0.7 + dataMix.code * 0.2)),
+    knowledge: clamp(rawCapability * (0.5 + dataMix.web * 0.3 + dataMix.books * 0.3)),
+    multimodal: clamp(rawCapability * (dataMix.images * 0.5 + dataMix.video * 0.4 + dataMix.audio * 0.2)),
+    agents: clamp(rawCapability * (0.2 + dataMix.code * 0.3 + dataMix.conversation * 0.2)),
+    speed: Math.max(1, 100 - architecture.totalParameters * 0.3 + efficiencyBonus * 2 + (architecture.type === 'moe' ? MOE_SPEED_MULTIPLIER * 10 : 0)),
+    contextUtilization: Math.min(100, architecture.contextWindow * 0.4),
   };
 
+  const breakthroughBonuses: Partial<Record<keyof ModelCapabilities, number>> = {};
+  for (const event of pipeline.events) {
+    if ((event.type === 'breakthrough' || event.type === 'emergent_capability') && event.impact.capabilityDomain && event.impact.capabilityBonus) {
+      const domain = event.impact.capabilityDomain;
+      breakthroughBonuses[domain] = (breakthroughBonuses[domain] ?? 0) + event.impact.capabilityBonus;
+    }
+  }
+  for (const [domain, bonus] of Object.entries(breakthroughBonuses)) {
+    const key = domain as keyof ModelCapabilities;
+    capabilities[key] = clamp(capabilities[key] + bonus);
+  }
+
+  const completedStages: ('pretraining' | 'sft' | 'alignment')[] = ['pretraining'];
+
+  if (pipeline.stages.sft?.isComplete) {
+    completedStages.push('sft');
+    const sft = pipeline.stages.sft;
+    for (let i = 0; i < sft.specializations.length; i++) {
+      const spec = sft.specializations[i];
+      const bonuses = SFT_SPECIALIZATION_BONUSES[spec];
+      if (!bonuses) continue;
+      const diminishing = i === 0 ? 1.0 : i === 1 ? 0.7 : 0.4;
+      for (const [cap, value] of Object.entries(bonuses)) {
+        const key = cap as keyof ModelCapabilities;
+        capabilities[key] = clamp(capabilities[key] + value * diminishing);
+      }
+    }
+  }
+
   const safetyResearch = state.research.completedResearch.filter(
     r => r.includes('alignment') || r.includes('interpretability') || r.includes('constitutional'),
   ).length;
-  const safetyScore = Math.min(100, 30 + safetyResearch * 15 + Math.random() * 10);
+  let overallSafety = Math.min(100, 30 + safetyResearch * 15 + Math.random() * 10);
+  let refusalRate = overallSafety > 60 ? 0.1 : 0.03;
 
-  const safetyPenalty = safetyScore > 60 ? (safetyScore - 60) * 0.1 : 0;
-  const benchmarkScore = Math.max(0,
-    (capabilities.reasoning * 0.3 + capabilities.coding * 0.25 +
-    capabilities.creative * 0.2 + capabilities.multimodal * 0.15 + capabilities.agents * 0.1) - safetyPenalty,
-  );
+  if (pipeline.stages.alignment?.isComplete) {
+    completedStages.push('alignment');
+    const alignment = pipeline.stages.alignment;
+    const methodConfig = ALIGNMENT_METHODS[alignment.method];
+    if (methodConfig) {
+      const safetyGain = methodConfig.safetyGain * alignment.safetyWeight;
+      overallSafety = Math.min(100, overallSafety + safetyGain);
+      refusalRate = methodConfig.baseRefusal * Math.pow(alignment.safetyWeight, 1.5);
+      const capLoss = methodConfig.capabilityLoss * alignment.safetyWeight * 0.5;
+      for (const key of Object.keys(capabilities) as (keyof ModelCapabilities)[]) {
+        if (key !== 'speed' && key !== 'contextUtilization') {
+          capabilities[key] = clamp(capabilities[key] - capLoss);
+        }
+      }
+    }
+  }
 
-  const parameterCount = Math.pow(10, generation) * (0.5 + Math.random());
+  const safetyProfile: SafetyProfile = {
+    overallSafety,
+    refusalRate,
+    harmAvoidance: overallSafety,
+    instructionFollowing: capabilities.reasoning * 0.8,
+    honesty: overallSafety * 0.9,
+  };
 
   return {
     id: uuid(),
-    name,
-    generation,
-    parameterCount,
-    trainingDataSize: dataTokens,
+    familyId: pipeline.familyId,
+    name: pipeline.modelName,
+    architecture,
+    dataMix,
     capabilities,
-    safetyScore,
-    benchmarkScore,
-    tuning: { preset: 'helpful-safe' },
+    safetyProfile,
+    rawCapability,
     isDeployed: false,
     trainedAtTick: state.meta.tickCount,
+    trainingCostTotal: compute,
+    trainingStagesCompleted: completedStages,
   };
 }
 
+function processVariantJobs(
+  state: GameState,
+  speedMultiplier: number,
+): { jobs: VariantCreationJob[]; newVariants: ModelVariant[] } {
+  const newVariants: ModelVariant[] = [];
+  const jobs = state.models.variantJobs.map(job => {
+    if (job.status !== 'active') return job;
+    const newProgress = job.progressTicks + speedMultiplier;
+    if (newProgress >= job.totalTicks) {
+      const baseModel = state.models.baseModels.find(m => m.id === job.baseModelId);
+      if (baseModel) {
+        const variant = createVariant(job, baseModel);
+        newVariants.push(variant);
+      }
+      return { ...job, status: 'completed' as const, progressTicks: job.totalTicks };
+    }
+    return { ...job, progressTicks: newProgress };
+  });
+  return { jobs, newVariants };
+}
+
+function createVariant(job: VariantCreationJob, base: BaseModel): ModelVariant {
+  const caps = { ...base.capabilities };
+  let costMultiplier = 1.0;
+  let speedMultiplier = 1.0;
+  let variantName = base.name;
+  let arch = { ...base.architecture };
+
+  if (job.jobType === 'distillation' && 'targetParameters' in job.config) {
+    const config = job.config;
+    const sizeRatio = config.targetParameters / base.architecture.totalParameters;
+    const retention = DISTILLATION_BASE_RETENTION + sizeRatio * 0.25;
+    for (const key of Object.keys(caps) as (keyof ModelCapabilities)[]) {
+      caps[key] = clamp(caps[key] * retention);
+    }
+    costMultiplier = sizeRatio * 0.8;
+    speedMultiplier = (1 / sizeRatio) * 0.7;
+    arch = { ...arch, totalParameters: config.targetParameters, activeParameters: config.targetParameters };
+    variantName = config.variantName;
+  } else if (job.jobType === 'fine-tuning' && 'specialization' in job.config) {
+    const config = job.config;
+    const bonuses = SFT_SPECIALIZATION_BONUSES[config.specialization];
+    if (bonuses) {
+      for (const [cap, value] of Object.entries(bonuses)) {
+        caps[cap as keyof ModelCapabilities] = clamp(caps[cap as keyof ModelCapabilities] + value);
+      }
+    }
+    variantName = config.variantName;
+  } else if (job.jobType === 'quantization' && 'level' in job.config) {
+    const config = job.config;
+    const qConfig = QUANTIZATION_CONFIGS[config.level];
+    if (qConfig) {
+      for (const key of Object.keys(caps) as (keyof ModelCapabilities)[]) {
+        if (key !== 'speed') caps[key] = clamp(caps[key] * qConfig.qualityRetention);
+      }
+      caps.speed = clamp(caps.speed * qConfig.speedMultiplier);
+      costMultiplier = qConfig.costMultiplier;
+      speedMultiplier = qConfig.speedMultiplier;
+    }
+    variantName = config.variantName;
+  }
+
+  return {
+    id: uuid(),
+    familyId: base.familyId,
+    baseModelId: base.id,
+    name: variantName,
+    variantType: job.jobType === 'distillation' ? 'distilled' : job.jobType === 'fine-tuning' ? 'fine-tuned' : 'quantized',
+    architecture: arch,
+    capabilities: caps,
+    safetyProfile: { ...base.safetyProfile },
+    isDeployed: false,
+    createdAtTick: 0,
+    quantization: job.jobType === 'quantization' && 'level' in job.config ? job.config.level : undefined,
+    distillationRetention: job.jobType === 'distillation' && 'targetParameters' in job.config
+      ? DISTILLATION_BASE_RETENTION + (job.config.targetParameters / base.architecture.totalParameters) * 0.25
+      : undefined,
+    finetuneSpecialization: job.jobType === 'fine-tuning' && 'specialization' in job.config ? job.config.specialization : undefined,
+    costMultiplier,
+    speedMultiplier,
+  };
+}
+
+function processEvalJobs(state: GameState): { jobs: EvalJob[]; newResults: BenchmarkResult[] } {
+  const newResults: BenchmarkResult[] = [];
+  const allModels: (BaseModel | ModelVariant)[] = [
+    ...state.models.baseModels,
+    ...state.models.families.flatMap(f => f.variants),
+  ];
+
+  const jobs = state.models.evalJobs.map(job => {
+    if (job.status !== 'active') return job;
+    const newProgress = job.progressTicks + 1;
+    if (newProgress >= job.totalTicks) {
+      const model = allModels.find(m => m.id === job.modelId);
+      if (model) {
+        const results = computeBenchmarkScores(model, job.benchmarkIds, state.meta.tickCount);
+        newResults.push(...results);
+        return { ...job, status: 'completed' as const, progressTicks: job.totalTicks, results };
+      }
+      return { ...job, status: 'completed' as const, progressTicks: job.totalTicks };
+    }
+    return { ...job, progressTicks: newProgress };
+  });
+  return { jobs, newResults };
+}
+
+function computeBenchmarkScores(
+  model: BaseModel | ModelVariant,
+  benchmarkIds: string[],
+  tick: number,
+): BenchmarkResult[] {
+  const benchmarkMap = new Map(BENCHMARKS.map(b => [b.id, b]));
+  return benchmarkIds.map(id => {
+    const bench = benchmarkMap.get(id);
+    if (!bench) return { benchmarkId: id, modelId: model.id, score: 0, ranAtTick: tick };
+    const primary = model.capabilities[bench.primaryCapability] ?? 0;
+    const secondary = bench.secondaryCapability ? (model.capabilities[bench.secondaryCapability] ?? 0) : 0;
+    const noise = (Math.random() - 0.5) * 6;
+    const score = clamp(primary * 0.7 + secondary * 0.3 + noise);
+    return { benchmarkId: id, modelId: model.id, score, ranAtTick: tick };
+  });
+}
+
+function computeVariantScore(variant: ModelVariant): number {
+  const c = variant.capabilities;
+  return (c.reasoning * 0.25 + c.coding * 0.2 + c.creative * 0.15 + c.math * 0.15 + c.knowledge * 0.15 + c.agents * 0.1);
+}
+
 function clamp(n: number): number {
   return Math.min(100, Math.max(0, n));
 }

packages/game-engine/src/systems/reputationSystem.ts

@@ -14,13 +14,9 @@ export interface ReputationTickResult {
 export function processReputation(state: GameState): ReputationState & { _safetyIncident?: boolean } {
   let { safetyRecord, publicPerception, employeeSatisfaction, regulatoryStanding } = state.reputation;
 
-  const bestModel = state.models.trainedModels
-    .filter(m => m.isDeployed)
-    .sort((a, b) => b.benchmarkScore - a.benchmarkScore)[0];
-
   let safetyIncident = false;
-  if (bestModel) {
-    const safetyLevel = bestModel.safetyScore;
+  if (state.models.bestDeployedSafetyScore > 0) {
+    const safetyLevel = state.models.bestDeployedSafetyScore;
     if (safetyLevel < LOW_SAFETY_THRESHOLD && state.meta.tickCount % 60 === 0) {
       const incidentProb = SAFETY_INCIDENT_PROBABILITY_BASE * (LOW_SAFETY_THRESHOLD - safetyLevel);
       if (Math.random() < incidentProb) {

packages/game-engine/src/tick.ts

@@ -40,13 +40,14 @@ export function processTick(state: GameState): Partial<GameState> {
   const stateWithInfra = { ...state, infrastructure };
   const modelResult = processModels(stateWithInfra);
 
-  if (modelResult.modelCompleted) {
+  for (const completed of modelResult.completedModels) {
     notifications.push({
       title: 'Training Complete',
-      message: `${modelResult.modelCompleted.name} is ready! Benchmark: ${modelResult.modelCompleted.benchmarkScore.toFixed(1)}/100`,
+      message: `${completed.name} is ready! Capability: ${completed.rawCapability.toFixed(1)}/100`,
       type: 'success',
     });
   }
+  notifications.push(...modelResult.notifications);
 
   const stateWithModels = { ...stateWithInfra, models: modelResult.modelsState };